ZGCM-1 packs heavy math and search skills into 7B parameters

ZGCM-1 is a new 7.39B-parameter dense model from Zhongguancun Academy and Zhongguancun Institute of Artificial Intelligence, built specifically for math reasoning and tool-assisted search.

It handles a 256K-token context window, supports both an internal thinking mode and direct responses in a single model, and can comfortably run on a standard consumer GPU. More importantly, the team open-sourced the entire training recipe from raw data pipelines to RL.

What it can do

Math reasoning is the headliner here. In thinking mode, ZGCM-1 scores 97.13% on MATH-500, 75.00% on AIME 2026, and 70.42% on HMMT 2025. Across 14 reasoning benchmarks evaluated against seven 7B to 8B models, it captured the top average rank.

It also handles active web retrieval and code exploration. On multi-step agent benchmarks, it hits 63.09% on WebWalkerQA, 19.43% on BrowseComp, and 42.52% on text-only GAIA. Using Ghidra reverse-engineering tools, it hit 62.00% on a binary function search benchmark, locating 31 out of 50 exact entry points.

How it works

Packing these capabilities into 7 billion parameters required a heavy focus on training efficiency and context management.

ZGCM-1 runs on a hybrid attention setup across 32 layers: 27 layers use gated sliding-window attention with a tight 128-token local window, while 5 layers handle global attention. This gave the team a 3.94x training throughput bump at 256K compared to standard full attention.

The training pipeline spans three distinct phases. Pretraining chewed through roughly 4.19T tokens using hybrid FP8 precision and the Muon optimizer, cutting pretraining time-to-loss by an estimated 4.2x. A 600B-token mid-training phase gradually stepped context up from 16K to 64K and 256K while framing tool interactions as Markov Decision Processes. Fine-tuning wrapped up with execution-verified trajectories and an assistant-only loss.

Specs and open recipe

At 7.39B parameters, ZGCM-1 fits easily on consumer hardware for local prototyping and private agent hosting.

The repository is licensed under MIT and gives you the exact setup instructions across five clean subdirectories: data processing, pretraining, mid-training, SFT, and reinforcement learning. Weights and training datasets are hosted directly on Hugging Face.

Why it matters

If you are an independent builder building agentic workflows or automated math grading, running a 70B model locally is usually out of the question. ZGCM-1 gives you high-end reasoning and tool-calling on hardware you already own.

Independent reviewers have pointed out that some baseline comparisons in the paper rely on older model checkpoints, which gives ZGCM-1 a slightly cherry-picked advantage on certain leaderboards. That said, getting 97% on MATH-500 alongside a fully disclosed, reproducible end-to-end training pipeline is a massive win for open research.